Hyperscaler capex hit a record $133B in Q1 (+70% y/y) but depreciation is compounding — consensus pegs annual depreciation for MSFT, AMZN, META, and GOOGL exceeding $430B within five years against $372B of combined net income last year. AI infrastructure demand is broadening past GPUs into power, optical, and packaging, while Huawei AI chip revenue surges toward $12B on forced domestic substitution.
Big Tech spent $133B in Q1 capex, up 70% y/y. Combined depreciation hit $41.6B in the quarter. Server equipment depreciates over 5-6 years — today's record spend locks in earnings drag through 2030. Consensus pegs annual depreciation for MSFT, AMZN, META, and GOOGL exceeding $430B within five years against $372B of combined net income last year. Useful lives already extended from ~4 years to 5-6. That lever is spent. Goldman's Covello recommends long hyperscalers / short semis, arguing the market has already rewarded chip stocks while hyperscaler valuation multiples compress on ROI skepticism.
↳ Counterpoint: Applied Digital just signed a 15-year, 300 MW lease with a second IG hyperscaler — adds $7.5B in contracted value, total now $23B+. Over 50% of revenue backed by IG tenants. Demand conviction at the infrastructure layer remains strong even as capex scrutiny intensifies.
Monolithic Power's Q1 confirms AI demand broadening into CPU servers, optical modules, switches, storage, DDR5, and higher-voltage rack-level power conversion. AXT reports indium phosphide substrate demand accelerating faster than qualified supply — customer procurement shifting to multi-year allocation security. Aixtron sizes the optoelectronics opportunity at 80-100 G10 tools per year (~€300-400M annual TAM). Epitaxial wafer ASPs are turning; leading-edge logic (≤7nm) wafer demand estimated to inflect next year, reaching ~1M wpm in CY28, ~10% of total 300mm equivalent demand. GlobalWafers, SUMCO, Shin-Etsu, Siltronics are the beneficiaries.
↳ The memory mix wrinkle: profitability in commodity DRAM/NAND now exceeds HBM for the Big 3, incentivizing commodity-first production. This tightens HBM supply exactly as accelerator demand accelerates. David Patterson flags High Bandwidth Flash (HBF) as the next bottleneck after HBM.
Huawei AI chip revenue projected at ~$12B in 2026 vs $7.5B in 2025, a 60%+ jump. The 950PR entered mass production in March; a 950DT upgrade is planned for Q4. Beijing is directing Chinese tech firms to support domestic suppliers and restrict NVIDIA chips to overseas operations. NVIDIA H200 chips reportedly cannot clear Chinese customs. Morgan Stanley forecasts China's AI chip market at $67B by 2030, with 86% from domestic suppliers. SMIC has produced early 3nm test wafers.
↳ Escalation risk: if the U.S. wants Gulf State cooperation on energy leverage against China, it may need to remove regional threats — implying further export control tightening ahead.
Preliminary April Korea semiconductor exports: DRAM $9.25B (+15% MoM, +83% QoQ), DRAM modules $6.14B (-17% MoM, +70% QoQ), NAND $1.67B (-34% MoM, +17% QoQ), SSD $3.84B (+20% MoM, +181% QoQ). The sequential snapback in DRAM and SSD is striking and suggests demand pull-in ahead of anticipated supply tightness.
Low TSMC SoIC yields are driving the industry to cut near-term 2026 CPO shipment forecasts. Former TSMC packaging VP Douglas Yu has joined MediaTek — insiders expect his presence to materially de-risk their first EMIB-T adoption. TSMC reportedly assumed customers wouldn't leave; they did. TSMC is now rushing to fill the packaging gap between CoWoS and alternatives. Teradyne beat Q1 ($1.28B revenue, 60.9% GM) but guided Q2 to $1.15-1.25B (midpoint -6% QoQ), signaling a 2H26 transition period.
GPT-55 reportedly reversing the competitive dynamic while Opus 4.7 underwhelms — attributed to Anthropic compute shortage versus OpenAI's compute lead. OpenAI has reportedly cracked recursive self-improvement for Codex. UK AISI cyber range evals show 5.5 performing similarly to Mythos on long-horizon agentic tasks. Inference throughput is emerging as the next battleground. Meanwhile Qualcomm is seeing an AI200-style bump on read-through from an unknown hyperscaler chip — mobile appears to have bottomed but memory remains "ultra unaffordable." Apple's Q2 confirms SoC/leading-edge capacity — not memory — is the binding constraint on unit shipments.
Full analysis, implications, relative trade read-throughs, positioning reads and catalysts — delivered by email at 07:00 HKT daily.